Container Images and CernVM-FS

CernVM-FS interacts with container technologies in two main ways:

  1. CernVM-FS application repositories (e.g. /cvmfs/ can be mounted into a stock container (e.g. CentOS 8)
  2. The container root filesystem (e.g. the root file system “/” of CentOS 8) itself can be served directly from CernVM-FS

Both ways have a similar goal, that is to give users access to a reproducible, ready-to-use environment while retaining the advantages of CernVM-FS regarding data distribution, content de-duplication, software preservation and ease of operations.

Mounting /cvmfs inside a container

The simplest way to access /cvmfs from inside a container is to bind-mount the /cvmfs host directory inside the container.

Using this approach will allow using small images to create a basic operating system environment, and to access all the necessary application software through /cvmfs.

This is supported by all the common containers runtimes, including:

  1. Docker
  2. Podman
  3. runc
  4. Singularity
  5. Kubernetes


To bind-mount CVMFS inside a docker container, it is sufficient to use the --volume/-v flag.

For instance:

docker run -it --volume /cvmfs:/cvmfs:shared ubuntu ls -lna /cvmfs/

Of course, it is also possible to limit the bind-mount to only one repository, or a few repositories:

$ docker run -it -v /cvmfs/ \
                 -v /cvmfs/ ubuntu
root@808d42605e97:/# ll /cvmfs/
total 17
drwxr-xr-x 17  125  130 4096 Nov 27  2012
drwxr-xr-x  8  125  130 4096 Oct 15  2018

Podman has the same interface as docker, but it requires the ro options when mounting a single repository.

$ podman run -it -v /cvmfs/ ubuntu ls -lna /cvmfs/
total 13
drwxr-xr-x  3     0     0 4096 Apr 20 11:34 .
drwxr-xr-x 22     0     0 4096 Apr 20 11:34 ..
drwxr-xr-x 17 65534 65534 4096 Nov 27  2012

A similar approach is possible with Singularity, but the syntax is a little different.

$ singularity exec --bind /cvmfs docker://library/ubuntu ls -l /cvmfs/
total 2
drwxrwxr-x.  3 cvmfs cvmfs  3 Jan  6  2011 etc
lrwxrwxrwx.  1 cvmfs cvmfs 16 Aug  6  2011 group_login.csh -> lib/etc/LHCb.csh
lrwxrwxrwx.  1 cvmfs cvmfs 15 Aug  6  2011 -> lib/etc/
drwxrwxr-x. 20 cvmfs cvmfs  3 Apr 24 12:39 lib

Also in singularity it is possible to use the syntax host_directory:container_directory and it is possible to mount multiple paths at the same time separating the --bind arguments with a comma.

$ singularity exec --bind /cvmfs/,/cvmfs/ \
    docker://library/ubuntu ls -l /cvmfs/
total 5
drwxr-xr-x 17      125      130 4096 Nov 27  2012
drwxrwxr-x  4      125      130    6 Nov 16  2010

For Kubernetes, the approach is more heterogeneous and it depends on the cluster settings.

For Kubernetes, a CSI-plugin makes it simple to mount a repository inside a Kubernetes managed container. The plugin is distributed and available to the CERN Kubernetes managed clusters.

Distributing container images on CernVM-FS

Image distribution on CernVM-FS works with _unpacked_ layers or image root file systems. Any CernVM-FS repository can store container images.

A number of images are already provided in /cvmfs/, a repository managed at CERN to host container images for various purposes and groups. The repository is managed using the DUCC utility.

Every container image is typically stored in two forms on CernVM-FS

  1. All the unpacked layers of the image
  2. The whole unpacked root filesystem of the image

Storing the layers of an image in CernVM-FS allows using (after creation) the docker thin images described in CernVM-FS Graph Driver Plugin for Docker, very small docker containers that compose the image’s filesystem from the layers stored in CernVM-FS. The docker thin image can be created using the DUCC utility.

If the whole filesystem of an image is stored in the repository it is possible to run the image using singularity:

singularity exec /cvmfs/\:centos7 /bin/bash


The repository provides a centrally managed container image hub without burdening users with managing their CernVM-FS repositories or conversion of images. It also enables saving storage space because of cvmfs deduplication of files that are common between different images. The repository is publicly available.

To add your image to you can add the image name to any one of the following two files.


The first file is accessible from CERN infrastructure, while the second is on Github open to everybody.

A simple pull request against one of those files is sufficient, the image is vetted, and the pull request merged. Soon after the pull request is merged DUCC publishes the image to /cvmfs/ Depending on the size of the image, ingesting an image in takes ~15 minutes.

The images are continuously checked for updates. If you push another version of the image with the same tag, DUCC updates the image on CVMFS, again with ~15 minutes of delay.

DUCC syntax for images

The image in DUCC must be specified following a simple format. The following examples are valid image specifications:

The first two refer to images in the classical docker hub, the standard centos using the latest tag and the cms version of centos8, again using the latest tag. The third image refers to a docker image hosted on CERN GitLab that contains the code for an analysis by a CERN user.

It is possible to use the * wildcard which acts like the * glob in the terminal shell to specify multiple tags.

For instance:*

is a valid image specification, and triggers conversion of all the atlas/analysisbase images whose tags start with 21.2.1, including:


But not:


Since it is 21. 3 .10 and not 21.2

The * wildcard can also be used to specify all the tags of an image, like in this example:*

All the tags of the image pyhf/pyhf that are published in docker hub will be published in

Updated images and new tags

DUCC polls the docker registries continuously. As soon as a new or modified container image is detected it starts the conversion process.

Work in progress

There are several lines of development that we are pursuing to improve the CernVM-FS container integration.

containerd remote-snapshotter plugin

This will allow running images from Kubernetes looking for the layers first in CernVM-FS and if the layers are not to be found, downloading them from the standard docker registry.

podman integration

Similarly to the containerd integration, this development will allow running a standard docker image using podman fetching the layers, unpacked, from a CernVM-FS repository, falling back to downloading the files from the registry if necessary.

DUCC registry interface

This development will allow for pushing the image to a special registry and for finding the image in the CernVM-FS repository as soon as the push finishes. While this will result in slower push operations since the layers need to be ingested into CernVM-FS, it will guarantee full distribution of the image as soon as the push completes.